Project Grant F32NS151141

EARLY IDENTIFICATION OF EPILEPSY SURGERY CANDIDATES USING LARGE LANGUAGE MODELS - PROJECT SUMMARY/ABSTRACT EPILEPSY AFFECTS ALMOST 3.5 MILLION PEOPLE IN THE UNITED STATES, AROUND ONE MILLION OF WHICH CONTINUE TO HAVE SEIZURES THAT DO NOT RESPOND TO ANTI-SEIZURE MEDICATIONS. DRUG-RESISTANT EPILEPSY (DRE) IS ASSOCIATED WITH POOR NEUROLOGIC OUTCOMES, QUALITY OF LIFE, DEPRESSION, AND CARRIES A 1% RISK OF SUDDEN UNEXPECTED DEATH IN EPILEPSY PER YEAR. AFTER FAILING AT LEAST TWO ADEQUATE DRUG TRIALS, THE ODDS THAT A PATIENT WITH DRE WILL OBTAIN LONG-TERM SEIZURE FREEDOM WITH ADDITIONAL DRUG TRIALS IS ONLY 8-10%. HOWEVER, RESECTIVE EPILEPSY SURGERY INCREASES THE CHANCES OF SEIZURE FREE TO 58-73% AND REDUCES THE RISK OF MORTALITY BY 66-81%. DESPITE CLASS 1 EVIDENCE AND CLINICAL PRACTICE GUIDELINES URGING EARLY REFERRAL FOR PRESURGICAL EVALUATION, THE AVERAGE SEIZURE DURATION AT THE TIME OF SURGERY IS 20 YEARS. SURVEYS SHOW THAT MANY PATIENTS ARE UNAWARE SURGICAL TREATMENT OPTIONS EXIST, AND CLINICIANS STRUGGLE TO SIFT THROUGH YEARS OF CLINICAL HISTORY AND TEST RESULTS PRIOR TO CLINIC VISITS. CLINICAL DECISION SUPPORT SYSTEMS CAN ASSIST CLINICIANS IN IDENTIFYING CANDIDATES FOR EPILEPSY SURGERY EARLIER IN THE DISEASE COURSE. A RANDOMIZED CONTROLLED TRIAL SHOWED ALERTS FROM AN ELECTRONIC HEALTH RECORD (EHR)-BASED SYSTEM INCREASED THE LIKELIHOOD THAT A PATIENT WAS REFERRED FOR A PRESURGICAL EVALUATION BY THREE TIMES. THIS AUTOMATED SURGICAL CANDIDATE IDENTIFICATION SYSTEM (ASCENT) USED TRADITIONAL MACHINE LEARNING AND NATURAL LANGUAGE PROCESSING METHODS TO PRODUCE PROBABILISTIC ESTIMATES THAT A PATIENT'S DATA WAS SIMILAR TO HISTORICAL SURGICAL PATIENTS' DATA. HOWEVER, IT LACKED ANY FUNDAMENTAL UNDERSTANDING OR THEORETICAL FRAMEWORK WHEN MAKING ITS PREDICTIONS. THIS MEANT THAT, DESPITE ITS RELATIVELY GOOD OVERALL PERFORMANCE (AUC = 0.91), ITS INDIVIDUAL PREDICTIONS WERE PLAGUED BY FALSE POSITIVES (PPV = 0.07). LARGE LANGUAGE MODELS (LLMS) OFFER A POTENTIAL SOLUTION. IN THE PROPOSED RESEARCH, I WILL FIRST DEVELOP AND PROSPECTIVELY VALIDATE A DOMAINADAPTED LLM (ASCENT-LLM) FOR IDENTIFYING EPILEPSY SURGERY CANDIDATES. I WILL UTILIZE AN EXPERTLY ANNOTATED CORPUS OF NEUROLOGY NOTES AND EEG AND MRI REPORTS TO FINE TUNE AN OPEN-SOURCE LLM. THIS WILL ENABLE ME TO SCREEN PATIENTS' EHRS TO INFORMATION NEEDED TO KNOW WHETHER THEY MEET CLINICAL CRITERIA FOR SURGICAL CANDIDATE, BASED ON ESTABLISHED GUIDELINES. SECOND, I WILL ANALYZE THE NATURAL HISTORY OF PATIENTS' TREATMENT COURSES OVER TIME TO UNDERSTAND HOW AND WHY DELAYS IN SURGICAL CARE ARE HAPPENING. THIRD, I WILL VALIDATE THE SYSTEM PROSPECTIVELY AND STUDY THE EFFECTS THAT ALERTING CLINICIANS AND PATIENTS HAS ON THE SUBSEQUENT LIKELIHOOD THAT THEY PURSUE SURGICAL INTERVENTIONS IN THE FUTURE. THIS PROPOSAL WILL PROVIDE ME WITH THE TRAINING NEEDED TO BECOME AND EXPERT IN LLM- BASED CARE INTERVENTIONS AND, IN THE FUTURE, SUBMIT K OR R GRANTS THAT WOULD IMPLEMENT ASCENT-LLM IN A MULTI- CENTER TRIAL.